Lessons Learned from Building Agentic Systems With Jayeeta Putatunda (Director of AI Center of Excellence at Fitch Group)

1
00:00:07,079 --> 00:00:07,769
Good morning.
2
00:00:07,979 --> 00:00:09,659
Uh, good afternoon everybody.
3
00:00:10,529 --> 00:00:13,829
Welcome to another, uh,
webinar on AI Explained.
4
00:00:13,879 --> 00:00:19,249
Today's topic is on Lessons Learned
from Building Agent Systems.
5
00:00:19,579 --> 00:00:20,639
I'm Krishna Gade.
6
00:00:20,689 --> 00:00:22,519
I'm the Founder and CEO of Fiddler AI.
7
00:00:23,149 --> 00:00:24,049
I'll be your host today.
8
00:00:24,520 --> 00:00:26,230
We have a very special guest today.
9
00:00:26,380 --> 00:00:31,840
Uh, Jayeeta Putatunda, a director of
AI Center of Excellence at Fitch Group.
10
00:00:32,700 --> 00:00:36,730
Jayeeta is a very, you know,
accomplished AI uh, leader.
11
00:00:36,760 --> 00:00:38,590
She's a rising star in this field.
12
00:00:38,950 --> 00:00:43,750
Her achievements include the AI100
Award in Generative AI, uh, recognition
13
00:00:43,750 --> 00:00:48,340
among Top 25 Visionary Women in
FinTech AI and selection as one of
14
00:00:48,340 --> 00:00:52,600
the 33 global advisors for NVIDIA's
Enterprise Platform Advisor program.
15
00:00:53,360 --> 00:00:57,070
Jayeeta, you know, please,
uh, join us on this program.
16
00:01:00,820 --> 00:01:01,600
Thank you very much.
17
00:01:01,600 --> 00:01:04,750
Thank you for the kind introduction,
and thank you for having me here today.
18
00:01:05,815 --> 00:01:06,175
Awesome.
19
00:01:06,355 --> 00:01:08,180
Excited to talk a bit about AI.
20
00:01:09,300 --> 00:01:09,580
Absolutely.
21
00:01:10,200 --> 00:01:14,095
Jayeeta, I think, you know, uh, you are
in the thick of things as in the field
22
00:01:14,095 --> 00:01:20,545
of AI, uh, uh, in generative AI now agent
systems and, uh, you know, I think as
23
00:01:20,545 --> 00:01:25,765
the world is moving in the world of AI
agents, um, what is the biggest reality
24
00:01:25,765 --> 00:01:31,510
check when you first moved your proof of
concept to a, a live production agent.
25
00:01:31,510 --> 00:01:35,170
You know, uh, if you can share any
interesting insights, that would be great.
26
00:01:35,740 --> 00:01:38,890
Yeah, I mean, that's a great point, right?
27
00:01:38,890 --> 00:01:41,980
Everybody and every now and then,
every day when I wake up, I see
28
00:01:41,980 --> 00:01:44,680
there are five different processes,
three different frameworks, and.
29
00:01:45,385 --> 00:01:48,445
When we are in this space and in
the tech, the goal is always to
30
00:01:48,445 --> 00:01:50,245
build something that that is new.
31
00:01:50,245 --> 00:01:52,405
That's better, that
works the best, but how?
32
00:01:52,465 --> 00:01:54,145
How do you really measure better?
33
00:01:54,175 --> 00:01:57,085
How do you really measure if
what you have in place right
34
00:01:57,085 --> 00:01:58,765
now, does it need an upgrade?
35
00:01:58,945 --> 00:02:01,555
If it needs an upgrade, what are
the baseline metrics that you're
36
00:02:01,555 --> 00:02:03,445
really evaluating it against?
37
00:02:03,655 --> 00:02:05,935
So my answer to that
would be twofold, right?
38
00:02:05,935 --> 00:02:09,445
First is, I think it's always been
there in the software development
39
00:02:09,445 --> 00:02:10,915
lifecycle, the 80 20 rule.
40
00:02:11,785 --> 00:02:15,385
The most of the time when, and I am
guilty of that too because I am a
41
00:02:15,385 --> 00:02:18,775
data scientist, uh, by profession and
I've been in the field for 10 years.
42
00:02:19,075 --> 00:02:22,645
My goal, I'm very, I get excited every
time I see a new model is like, I have
43
00:02:22,645 --> 00:02:26,035
to implement that and implement that with
a new framework and see how it works.
44
00:02:26,335 --> 00:02:30,295
But I think as in when you think more
from not only like your developer
45
00:02:30,295 --> 00:02:33,355
perspective, but more of like the
business function that how are you
46
00:02:33,355 --> 00:02:35,185
really adding value to the business?
47
00:02:35,395 --> 00:02:37,345
Is the, there is the ROI worth it.
48
00:02:37,345 --> 00:02:39,415
You kind of do that 80 20 rule again.
49
00:02:39,745 --> 00:02:45,520
80% of the focus should be on, uh,
all those use cases where like they
50
00:02:45,520 --> 00:02:48,760
say low hanging fruit, but that
doesn't mean that it's not important.
51
00:02:48,760 --> 00:02:55,060
It means that it has a bigger impact and
your impact to effort ratio is like good.
52
00:02:55,120 --> 00:02:58,060
And you can't take that on rather
than spending, I don't know,
53
00:02:58,060 --> 00:02:59,440
six months just prototyping.
54
00:02:59,440 --> 00:03:02,530
And by that time the framework has
already been obsoleted and we've been
55
00:03:02,530 --> 00:03:03,970
seeing it just talking about it, right?
56
00:03:03,970 --> 00:03:06,550
Like some models came up six
months ago and now they're
57
00:03:06,550 --> 00:03:08,020
like extinct from the scene.
58
00:03:08,020 --> 00:03:10,690
And we are like, okay, what
are we really doing here?
59
00:03:11,130 --> 00:03:12,630
So that's my first take.
60
00:03:12,630 --> 00:03:16,600
And I think the, the second biggest
lesson I, I feel is that most of the
61
00:03:16,600 --> 00:03:19,960
time, and it has been true even during
like when we were building predictive
62
00:03:19,960 --> 00:03:23,800
ml machine learning models, but it was
easier to handle then because our systems
63
00:03:23,800 --> 00:03:27,550
were deterministic and we had like a
good set of historical data that we
64
00:03:27,550 --> 00:03:31,360
were comparing it against and comparing
one is to like, like apples to apples.
65
00:03:31,660 --> 00:03:35,455
Now it's more like the generative AI
systems are giving so much output.
66
00:03:35,455 --> 00:03:37,105
How do you really categorize it?
67
00:03:37,105 --> 00:03:40,375
How do you really, uh, define your matrix?
68
00:03:40,465 --> 00:03:43,735
And not just matrix that, oh,
it gave me productivity gains.
69
00:03:43,765 --> 00:03:44,905
What productivity gains?
70
00:03:44,905 --> 00:03:47,395
Is it, say how saving, how
much of your dollars, how
71
00:03:47,395 --> 00:03:49,495
much of your developer's time?
72
00:03:49,495 --> 00:03:54,545
Or is it like adding, um, I don't know,
like reducing cycle of completion.
73
00:03:54,725 --> 00:04:00,305
So very specific methods of evaluating of
why you are really doing an agentic AI,
74
00:04:00,365 --> 00:04:02,555
um, application or building it through.
75
00:04:02,945 --> 00:04:07,055
But yeah, I feel like if these two
processes are nailed at the head before
76
00:04:07,055 --> 00:04:11,495
you begin your process, the chances of
success definitely increases multifold.
77
00:04:12,445 --> 00:04:12,985
Absolutely.
78
00:04:13,195 --> 00:04:16,375
So before we dive in deeper, right,
can you give us an overview of the
79
00:04:16,375 --> 00:04:19,825
type of agent tech systems and products
you have built recently and deployed?
80
00:04:19,825 --> 00:04:19,915
Sure.
81
00:04:20,185 --> 00:04:21,625
What are some of the use cases?
82
00:04:21,625 --> 00:04:23,215
What are these systems trying to solve?
83
00:04:23,605 --> 00:04:27,955
Yeah, I mean there are, and
so a couple of things, right?
84
00:04:28,255 --> 00:04:32,815
There are two ways of defining what
kind of agent solutions you would
85
00:04:32,815 --> 00:04:34,525
build first based on your use cases.
86
00:04:34,525 --> 00:04:37,375
And second, based on the agentic
AI frameworks that's available.
87
00:04:37,735 --> 00:04:41,545
Uh, by frameworks I mean like, is it
going to be a simple react model where
88
00:04:41,545 --> 00:04:45,325
it's like reasoning and doing some action
for you, a one-to-one output, or it's
89
00:04:45,325 --> 00:04:49,735
going to be more of a reflection based
model where you're also auto optimizing
90
00:04:49,735 --> 00:04:54,115
on the spot, uh, of, of whatever the
output you are, uh, you know, outputting
91
00:04:54,295 --> 00:04:57,505
based on some business rules or like
some evaluation criteria that you
92
00:04:57,505 --> 00:04:59,785
already have from the, the business side.
93
00:05:00,025 --> 00:05:04,135
The second part is I think we also
have to figure out that where is
94
00:05:04,135 --> 00:05:08,995
that high ROI agentic pattern,
like what are the use cases?
95
00:05:08,995 --> 00:05:09,655
Some use cases.
96
00:05:09,655 --> 00:05:13,765
And especially in finance, I still
believe that we are not at that space
97
00:05:13,765 --> 00:05:17,845
where agents can really, and this is my
personal opinion, uh uh, based on like
98
00:05:17,845 --> 00:05:21,385
what I'm seeing in the industry overall,
the kinds of data that I'm handling.
99
00:05:21,535 --> 00:05:25,015
Is that the highest autonomy use cases or?
100
00:05:25,555 --> 00:05:27,985
The patterns is, we
are not ready for that.
101
00:05:28,285 --> 00:05:31,855
I don't think we will ever be ready
because we are also very highly
102
00:05:32,125 --> 00:05:33,985
govern and it's, it's for the good.
103
00:05:33,985 --> 00:05:38,155
I feel like there has to be some layer
of credibility and responsibility,
104
00:05:38,155 --> 00:05:39,805
accountability that needs to happen.
105
00:05:39,835 --> 00:05:43,225
Uh, and, uh, the entire autonomy pipeline,
I don't think we are gonna go there.
106
00:05:43,465 --> 00:05:47,410
So how do you, how do you find a real
balance and find some use cases that that
107
00:05:47,410 --> 00:05:52,660
is really, uh, giving you a lot of, uh,
time savings and your analysts are freeing
108
00:05:52,660 --> 00:05:56,860
up their time to do some more high value
analysis versus, I don't know, like going
109
00:05:56,860 --> 00:06:03,010
and reading a 500 uh, page PDF, like
for and spending like three hours on it.
110
00:06:03,010 --> 00:06:07,270
Rather than if you can build a system or
like an agent, TKI pipeline where you are
111
00:06:07,270 --> 00:06:11,080
processing the data the right way, you
have a interface for them to really query
112
00:06:11,560 --> 00:06:15,820
conversation and then find exact places
of information that they're looking for.
113
00:06:15,820 --> 00:06:16,990
That's much high value.
114
00:06:17,320 --> 00:06:17,950
So I would say.
115
00:06:18,180 --> 00:06:19,680
Yeah, use case driven.
116
00:06:19,800 --> 00:06:24,510
Um, uh, uh, some use cases that definitely
comes into mind is like maybe report
117
00:06:24,510 --> 00:06:29,820
generations having some kind of voice
and templatized, uh, uh, models ready
118
00:06:29,820 --> 00:06:33,240
and then utilizing some of these agentic
AI frameworks to see that, how do you
119
00:06:33,240 --> 00:06:37,890
really hone in on the, uh, optimization
side and output something that's
120
00:06:37,980 --> 00:06:39,750
really aligns to what business wants.
121
00:06:39,900 --> 00:06:42,960
Again, at the end of the day, whatever
we are building or whatever anybody
122
00:06:42,960 --> 00:06:46,260
should build should align to what
the business requirements is and
123
00:06:46,545 --> 00:06:48,000
not, not just for the sake of it.
124
00:06:49,350 --> 00:06:52,020
So, so you mentioned two different
types of agent tech systems, right?
125
00:06:52,020 --> 00:06:55,320
One is more like a workflow like
system where maybe somewhat of
126
00:06:55,320 --> 00:06:56,670
determinism is baked into it.
127
00:06:57,020 --> 00:07:00,330
And then there's like a fully autonomous
agent that uses reflection and all.
128
00:07:00,570 --> 00:07:04,950
Can you, can you sort of go a little
deeper into these two systems where you
129
00:07:04,950 --> 00:07:09,060
have found one better or the other then
maybe get a little bit of internal yeah?
130
00:07:09,510 --> 00:07:10,050
Absolutely.
131
00:07:10,080 --> 00:07:11,850
Uh, again, like without.
132
00:07:12,865 --> 00:07:17,425
Going too deep or technical or like,
you know, revealing a lot of the PIs.
133
00:07:17,425 --> 00:07:21,865
I think one of the biggest area where
I feel a little process oriented
134
00:07:21,865 --> 00:07:26,395
workflow works is when, think
about some RPA processes, right?
135
00:07:26,395 --> 00:07:30,895
Or any maybe the software development
processes where we, where you were
136
00:07:30,895 --> 00:07:34,495
maybe, uh, calling three different
APIs, trying to gather the data
137
00:07:34,495 --> 00:07:37,465
and then do some processing and
then output in some certain way.
138
00:07:37,465 --> 00:07:41,305
We have all had those simpler, simpler
tasks or simpler processes, right?
139
00:07:41,820 --> 00:07:48,450
Those processes, I, I do not think really
needs a lot of, um, porosity from the
140
00:07:48,450 --> 00:07:50,790
models that the genAI usually adds.
141
00:07:50,790 --> 00:07:52,380
And it adds a lot of complexity, right?
142
00:07:52,380 --> 00:07:55,950
Like, why would I want to complicate
my clean workflow that I had, but
143
00:07:55,950 --> 00:07:57,120
I still want to make it better.
144
00:07:57,120 --> 00:08:01,435
I want to still give it the
option to call additional tools or
145
00:08:01,435 --> 00:08:05,305
additional memory just in case, so
that it remembers the conversations
146
00:08:05,305 --> 00:08:07,375
prior or whatever that memory is.
147
00:08:07,375 --> 00:08:07,645
Right?
148
00:08:07,855 --> 00:08:11,215
So maybe that's a very good use case
where you're following a workflow,
149
00:08:11,215 --> 00:08:15,835
but you're also giving a little bit
more flexibility in terms of tool, ac,
150
00:08:16,195 --> 00:08:20,165
access to memory, access to different
databases, and and then giving the
151
00:08:20,185 --> 00:08:24,770
user an end-to-end a better experience
of working through that workflow.
152
00:08:25,010 --> 00:08:28,160
It's a new workflow for the user,
but still with a little, I would say,
153
00:08:28,550 --> 00:08:34,370
augmented, um, features and capabilities
to really make their life better first.
154
00:08:34,370 --> 00:08:38,090
And the second one, of course, like
we said, uh, that is a little bit more
155
00:08:38,540 --> 00:08:44,660
where you are really trying to give the
model or whatever the agentic system say.
156
00:08:44,750 --> 00:08:48,590
Say you're creating a batch of three
systems and you want it to configure or
157
00:08:48,590 --> 00:08:53,275
like coordinate between each other and see
what the output of the previous agent was.
158
00:08:53,485 --> 00:08:57,055
Say if it was an evaluation agent
and you have a reflection agent, the
159
00:08:57,055 --> 00:09:00,115
reflection agent is supposed to take
the output of the evaluation agent
160
00:09:00,175 --> 00:09:03,865
and then go and clarify with all the
previous business rules you have inputted
161
00:09:03,865 --> 00:09:05,725
and see that, does this make sense?
162
00:09:05,785 --> 00:09:07,075
Are we avoiding by the rules?
163
00:09:07,135 --> 00:09:12,025
Did the output, uh, really follow all
the, uh, processes, score it on that,
164
00:09:12,025 --> 00:09:16,105
and then feed it back again to make sure
you're auto optimizing on, on the output.
165
00:09:16,105 --> 00:09:20,795
So here you have a little bit more
autonomy, but again, that autonomy is
166
00:09:20,795 --> 00:09:25,475
driven by the set of rules that came from
business and you are not just evaluating
167
00:09:26,015 --> 00:09:27,725
blindly if, if that makes sense.
168
00:09:27,725 --> 00:09:27,815
Right.
169
00:09:27,995 --> 00:09:31,595
So it seems like in the first case, you
are constructing the workflow manually,
170
00:09:31,715 --> 00:09:36,425
so the routing is almost deterministic,
but you are taking advantage of the model
171
00:09:36,425 --> 00:09:39,605
calls and tool calls to augment your
existing business process automation.
172
00:09:39,905 --> 00:09:39,995
Yep.
173
00:09:39,995 --> 00:09:44,365
In the second case, it seems like the
agent itself automates their workflow
174
00:09:44,365 --> 00:09:49,435
and it is, it is using reflection and
evaluation to self-correct itself, right?
175
00:09:49,795 --> 00:09:50,065
Yeah.
176
00:09:50,065 --> 00:09:53,965
I mean, it's planning a little bit by
itself and seeing that it, at this stage,
177
00:09:53,965 --> 00:09:58,255
do I still need to optimize or am I good
to go and just end the process and give
178
00:09:58,255 --> 00:10:02,050
the user the output that that they're
waiting for, so that, that decisioning
179
00:10:02,050 --> 00:10:04,900
is still happening from the agent
side and that's why I say maybe it's a
180
00:10:04,900 --> 00:10:08,570
little bit more autonomous, but again,
everything is like checked end-to-end.
181
00:10:08,590 --> 00:10:12,670
So the, it's not really autonomous,
it's just like, uh, uh, uh,
182
00:10:12,850 --> 00:10:14,140
somewhere, somewhere in the middle.
183
00:10:14,140 --> 00:10:14,830
Somewhere up there.
184
00:10:15,880 --> 00:10:16,030
Yeah.
185
00:10:16,030 --> 00:10:19,295
So, so in the second case in the
autonomous agents, this concept of
186
00:10:19,295 --> 00:10:22,120
reflection is quite interesting, right?
187
00:10:22,120 --> 00:10:25,930
Because, uh, in traditional software,
you, you don't have this, could
188
00:10:25,930 --> 00:10:29,630
you elaborate on what reflection is
and how it helps autonomous agents.
189
00:10:30,020 --> 00:10:30,680
Absolutely.
190
00:10:30,950 --> 00:10:36,110
So I, I, I forget where I read,
I was reading some, uh, uh, blog
191
00:10:36,110 --> 00:10:37,400
and I can share the link later.
192
00:10:37,400 --> 00:10:40,100
Is that how software is changing?
193
00:10:40,100 --> 00:10:40,310
Right?
194
00:10:40,310 --> 00:10:43,790
Like how the software development,
like lifecycle as we know is
195
00:10:43,790 --> 00:10:46,130
like changing a bit because.
196
00:10:46,540 --> 00:10:49,420
Yeah, like you said, we were
not using LLMs prior to this.
197
00:10:49,690 --> 00:10:51,430
LLMs are more non-deterministic.
198
00:10:51,430 --> 00:10:55,210
So how, what happens when you
bring in non-determinism into a
199
00:10:55,210 --> 00:10:56,830
deterministic workflow pipeline?
200
00:10:57,040 --> 00:10:59,440
How many checkpoints
do you need to measure?
201
00:10:59,590 --> 00:11:01,750
What do you need to measure
for each checkpoints?
202
00:11:01,870 --> 00:11:05,080
And also, like, not really
over-engineer it, otherwise it's
203
00:11:05,080 --> 00:11:07,870
gonna be like a too complicated system
with five different agents where
204
00:11:07,870 --> 00:11:09,520
you really don't need that many.
205
00:11:09,790 --> 00:11:13,140
Uh, maybe it can just be one trigger
agent and the rest of the workflow
206
00:11:13,140 --> 00:11:16,320
remains as is with the additional
capability of tools, like I was saying.
207
00:11:16,620 --> 00:11:22,230
So when, when you use something like
reflection, it's, it's mostly for
208
00:11:22,290 --> 00:11:24,930
the LLM to really critique itself.
209
00:11:24,930 --> 00:11:30,360
So underlying concept is still
LLM-as-a-judge, but again, being like
210
00:11:30,360 --> 00:11:36,280
responsible builders, like as like
consider myself to be, I really don't want
211
00:11:36,280 --> 00:11:41,260
to give all the autonomy or the decision
making to my reflection agent itself.
212
00:11:41,290 --> 00:11:46,540
Because I think there was some studies
also where, uh, we saw, and I, I know
213
00:11:46,750 --> 00:11:49,420
I'm, again, I'm, there's too many
papers that's going around, right.
214
00:11:49,420 --> 00:11:52,450
I completely missed the headlines,
but it's basically talked about how
215
00:11:52,760 --> 00:11:58,400
uh, uh, uh, LLM is a little bit biased
towards output from another LLM and
216
00:11:58,400 --> 00:12:02,060
it can like figure out and then say,
oh, this is better than something
217
00:12:02,060 --> 00:12:05,090
very similar, but written by a human
agent, maybe in a different way.
218
00:12:05,300 --> 00:12:08,870
So how do you, how do we take
all of these into consideration
219
00:12:08,870 --> 00:12:10,340
when we are building that system?
220
00:12:10,610 --> 00:12:14,570
So we have smaller checkpoints
also, we have very specific business
221
00:12:14,570 --> 00:12:17,570
guidelines that the LLM is using.
222
00:12:18,175 --> 00:12:22,405
From the historical data as well as
the current, uh, workflow data that the
223
00:12:22,405 --> 00:12:27,835
reflection agent is evaluating alongside
the output from the previous, uh, eval
224
00:12:27,835 --> 00:12:29,095
component that we were talking about.
225
00:12:29,095 --> 00:12:29,365
Right?
226
00:12:29,845 --> 00:12:29,935
Yep.
227
00:12:29,935 --> 00:12:35,605
So again, I don't think there is
like a, um, right way to do it.
228
00:12:35,635 --> 00:12:38,815
It's a little bit of trial
error to see what works with
229
00:12:38,900 --> 00:12:40,075
the use case you're handling.
230
00:12:41,230 --> 00:12:44,770
All with the mindset of really making
sure that you are, you are building a
231
00:12:44,770 --> 00:12:49,300
responsible system that should not be
biased towards any particular system.
232
00:12:49,300 --> 00:12:53,560
And the truth should be, uh,
yeah, following end to end process
233
00:12:53,560 --> 00:12:55,060
workflow, the best way it can.
234
00:12:55,660 --> 00:12:58,375
So it seems like self-reflection
sometimes could be biased.
235
00:12:59,140 --> 00:13:03,070
And so a third party evaluation,
third party reflection could be
236
00:13:03,100 --> 00:13:06,490
interesting and where it can encode
all the business rules and, and sort
237
00:13:06,490 --> 00:13:07,990
of, you know, judge the system, right?
238
00:13:08,230 --> 00:13:12,160
So that, that sort of brings down to like,
how do we test these agentic systems?
239
00:13:12,160 --> 00:13:16,030
Because in traditional software you
have deterministic inputs and outputs.
240
00:13:16,030 --> 00:13:17,890
You can test for this,
you can write all the.
241
00:13:18,220 --> 00:13:20,330
Your TD like test and development.
242
00:13:20,750 --> 00:13:24,040
But as you said, agent tech systems
are inherently non-deterministic.
243
00:13:24,490 --> 00:13:24,670
Mm-hmm.
244
00:13:24,880 --> 00:13:28,600
Have you, you know, sort of thought
about this approach of, you know, testing
245
00:13:28,600 --> 00:13:30,070
and validating this, these systems?
246
00:13:30,190 --> 00:13:34,060
Yeah, so, so I, I think it
has to be done in stages.
247
00:13:34,060 --> 00:13:35,140
Like it doesn't matter.
248
00:13:35,140 --> 00:13:38,710
And since we are, again, like I
said, highly goverance, matrix, and
249
00:13:39,430 --> 00:13:44,320
observability is a big thing that we
internally do like with a lot of priority
250
00:13:44,320 --> 00:13:48,850
and a lot of focus, and that is something
also that I really appreciate, but how
251
00:13:48,850 --> 00:13:52,360
do you really get started to make sure
your eval components will be ready?
252
00:13:52,540 --> 00:13:56,710
You have to have that, I call it the
data prep tax because if you're, I
253
00:13:56,710 --> 00:14:00,205
call it the tax because it's a lot
of, again, data data preparation
254
00:14:00,205 --> 00:14:04,255
issue making data, like they say,
AI ready, it's everywhere, right?
255
00:14:04,255 --> 00:14:06,925
Everybody's talking about how do
you really make your data AI ready.
256
00:14:06,925 --> 00:14:11,635
That means that using that data for
building your systems as well as keeping
257
00:14:11,635 --> 00:14:14,125
that data ready to evaluate your systems.
258
00:14:14,395 --> 00:14:18,505
So there's no other,
there's no magic pill here.
259
00:14:18,565 --> 00:14:22,705
I think data prep tax, like
we would like to call it, uh,
260
00:14:22,765 --> 00:14:24,325
really needs to be the focus for.
261
00:14:25,025 --> 00:14:28,805
Any business or I think it's more
important for like bigger and legacy
262
00:14:28,805 --> 00:14:33,305
systems where data is like really
like unstructured or unstructured,
263
00:14:33,305 --> 00:14:35,285
meaning it's in very different places.
264
00:14:35,495 --> 00:14:38,225
You need to bring it together,
make sure they align together.
265
00:14:38,285 --> 00:14:41,225
They have some kind of
lineage as well as versioning.
266
00:14:41,255 --> 00:14:43,655
And that is how you really track.
267
00:14:44,075 --> 00:14:48,635
Uh, I, I, uh, I was speaking to somebody,
um, in our like industry and I think
268
00:14:48,635 --> 00:14:55,090
this, this concept of versioning, like
how do you really version prompts as
269
00:14:55,090 --> 00:14:56,980
well as how do you really version eval.
270
00:14:57,370 --> 00:15:00,850
So eval outputs, so treat
it as like, say API.
271
00:15:01,030 --> 00:15:05,530
How do you really, uh, when you use API,
you version it, you make sure that with
272
00:15:05,530 --> 00:15:09,280
every upgraded version you have like
a different set of test cases ready.
273
00:15:09,550 --> 00:15:13,960
Similar for like anything that goes into
your system, maybe prompts the data,
274
00:15:13,960 --> 00:15:16,180
the system prompts the business rules.
275
00:15:16,425 --> 00:15:21,555
The voice or style prompts, all of that
needs to be versioned and then output
276
00:15:21,555 --> 00:15:23,535
it against, uh, how is it changing?
277
00:15:23,535 --> 00:15:27,615
Depending on the model, depending on if
you add an external tool, if you add an
278
00:15:27,615 --> 00:15:33,465
external step in reflection, and that has
to be like an end to end view for you or
279
00:15:33,465 --> 00:15:38,420
like the developer, uh, and the business
leaders to kind of really, uh, end to end.
280
00:15:39,685 --> 00:15:43,555
You know, we, you know, we all, coming
from data science ML background, we
281
00:15:43,555 --> 00:15:45,805
have seen how to evaluate ML models.
282
00:15:45,805 --> 00:15:50,125
You can create confusion matrices,
measure precision recall accuracy scores.
283
00:15:50,125 --> 00:15:50,425
Right.
284
00:15:50,785 --> 00:15:54,745
Now as we move to the world of
agent systems and agent AI, how
285
00:15:54,745 --> 00:15:56,275
does evaluation change here?
286
00:15:56,275 --> 00:15:59,515
You know, what do you, how do you
need to evaluate these agent systems?
287
00:15:59,515 --> 00:16:02,575
Can you give brief overview
of all the type of metrics and
288
00:16:02,575 --> 00:16:04,075
things that one needs to measure?
289
00:16:04,130 --> 00:16:04,855
Yeah, yeah.
290
00:16:05,305 --> 00:16:06,115
Uh, absolutely.
291
00:16:06,115 --> 00:16:06,625
So.
292
00:16:07,645 --> 00:16:09,955
Again, I would like to break it
down into four or five different
293
00:16:09,955 --> 00:16:14,350
components because I don't think one,
there's one magic evaluation, uh,
294
00:16:14,355 --> 00:16:16,075
uh, metric that you need to track.
295
00:16:16,465 --> 00:16:17,125
Couple, right?
296
00:16:17,125 --> 00:16:18,745
Like, like I said, traceability.
297
00:16:18,835 --> 00:16:21,325
Do you really have enough
logging matrix for all your.
298
00:16:22,210 --> 00:16:26,350
Calls all your tool calls if the output
of the tool calls were really correct.
299
00:16:26,470 --> 00:16:29,860
Uh, is there a way that you
can, uh, compare the links?
300
00:16:29,950 --> 00:16:33,370
Like say if you're doing a deep research
agent in one of your step and it's
301
00:16:33,430 --> 00:16:37,510
trying to find some links from the web
and it suggested the user links, is
302
00:16:37,510 --> 00:16:42,010
there a way for the system to evaluate
if the links are correct as well as
303
00:16:42,010 --> 00:16:47,095
like top rated, uh, and not like some,
you know garbage or like low quality,
304
00:16:47,395 --> 00:16:49,045
uh, links that came up in the search.
305
00:16:49,075 --> 00:16:51,355
So that's more for the
traceability and logging.
306
00:16:51,595 --> 00:16:54,685
Second is like, how are you
really using the models?
307
00:16:54,685 --> 00:16:59,395
Because again, models doesn't matter
how cheap it gets, at the end of
308
00:16:59,395 --> 00:17:00,775
the day, it starts compounding.
309
00:17:00,775 --> 00:17:03,925
When you are building a complicated
agenting system, it's not.
310
00:17:04,130 --> 00:17:05,720
About only one model call.
311
00:17:05,720 --> 00:17:10,220
It's about multimodal multi-model
calls in different layers and
312
00:17:10,220 --> 00:17:11,360
sometimes at the same time.
313
00:17:11,360 --> 00:17:14,780
Because if you're initiating like a,
say three different agents, it's calling
314
00:17:14,780 --> 00:17:18,410
two different tools you are going
or routing that through some uh, uh,
315
00:17:18,470 --> 00:17:20,210
models to really get the output of it.
316
00:17:20,480 --> 00:17:21,195
So how do you do track?
317
00:17:21,835 --> 00:17:24,565
Like track the token usages response time.
318
00:17:24,565 --> 00:17:25,615
Were there error rates?
319
00:17:25,855 --> 00:17:30,505
How many time your, uh, processes failed
because of a modeling issue, or is it
320
00:17:30,505 --> 00:17:34,435
because of like, it failed to generate
a response out out of that issue?
321
00:17:34,645 --> 00:17:35,815
Both are important, right?
322
00:17:35,815 --> 00:17:39,760
Like otherwise you are not setting up
your system to be like a really good.
323
00:17:40,370 --> 00:17:44,990
User experience for your customers or
whoever you are, um, opening that up to.
324
00:17:44,990 --> 00:17:49,820
So as well as like, there's ton, tons of
others, like how do we do drift detection?
325
00:17:49,820 --> 00:17:54,440
And then all that still maintains, but
there are too many other components that's
326
00:17:54,440 --> 00:17:58,730
now equally important, especially from
the infrastructure, uh, maintainability
327
00:17:58,730 --> 00:18:00,470
and infrastructure Observability as well.
328
00:18:01,280 --> 00:18:01,700
Awesome.
329
00:18:02,000 --> 00:18:05,480
And so when you diagnose these failures,
right, how do you, you know, what do
330
00:18:05,480 --> 00:18:09,680
you need to, like trace through or
debug the system failure so that you.
331
00:18:10,015 --> 00:18:12,835
Can understand if things are
happening because of tool calls
332
00:18:12,835 --> 00:18:14,230
or like model, model issues.
333
00:18:14,980 --> 00:18:17,830
Yeah, I think, uh, I think I, I
don't have the right answer to it,
334
00:18:17,830 --> 00:18:21,520
but it's, it's definitely work in
progress and what I have seen working
335
00:18:21,520 --> 00:18:24,250
is having checkpoints at every point.
336
00:18:24,250 --> 00:18:28,570
Like I said, like if you're building one
agent and that agent has five different
337
00:18:28,570 --> 00:18:33,700
steps that it's supposed to take, supposed
to take after each stage, there has to
338
00:18:33,700 --> 00:18:36,790
be like some component of logging that
it is doing that, okay, this was my
339
00:18:36,790 --> 00:18:40,810
input, this was my output, this is what
I called, and this was the response.
340
00:18:40,840 --> 00:18:41,565
Sometimes it might.
341
00:18:42,565 --> 00:18:46,045
Get too much of an information to
track, but it's worth the work,
342
00:18:46,105 --> 00:18:48,715
at least initially when you're
setting up for your use case.
343
00:18:49,135 --> 00:18:53,275
As you become mature and you get to
like understand a little bit of the
344
00:18:53,275 --> 00:18:56,305
workflow that you're building for,
you can tone it down a little bit,
345
00:18:56,395 --> 00:18:57,595
but I think at the end of the day.
346
00:18:58,015 --> 00:19:00,865
The goal, like nobody said
too much data is harmful.
347
00:19:00,865 --> 00:19:03,085
Like everybody said,
scars of data is harmful.
348
00:19:03,085 --> 00:19:05,485
So logging too much is my way to go.
349
00:19:06,550 --> 00:19:10,195
I, I log as much as possible even
sometimes when it's not required.
350
00:19:10,195 --> 00:19:12,685
But you never know when you can
find some, you know, gold mine
351
00:19:12,685 --> 00:19:17,335
an um, idea or like some kind of
thing that you're not thinking
352
00:19:17,335 --> 00:19:18,835
through, uh, from the data you log.
353
00:19:19,870 --> 00:19:20,230
Got it.
354
00:19:20,380 --> 00:19:23,110
And where does like
observability fit in here?
355
00:19:23,140 --> 00:19:26,950
Because, uh, you know, traditionally
observability, you know, has to deal
356
00:19:26,950 --> 00:19:30,700
with, you know, more deterministic
software where you are measuring
357
00:19:30,700 --> 00:19:33,850
reliability, latency, throughput Yeah.
358
00:19:33,850 --> 00:19:35,380
You know, server utilization, right?
359
00:19:35,710 --> 00:19:37,690
What, what do you think
about observability in
360
00:19:37,690 --> 00:19:38,470
the context of agentic AI?
361
00:19:39,580 --> 00:19:42,850
Yeah, I think that
definition still holds true.
362
00:19:42,940 --> 00:19:45,280
I will, I will not take it away at all.
363
00:19:45,670 --> 00:19:47,680
Uh, but there are additional.
364
00:19:48,550 --> 00:19:51,490
Additional angles to where you
would define observability.
365
00:19:52,570 --> 00:19:56,770
One thing that I really feel, um,
excited about is that, at least
366
00:19:56,775 --> 00:20:00,520
in, in the finance industry as much
as I've seen observability is no
367
00:20:00,520 --> 00:20:03,880
longer like a afterthought, like
it used to happen in most of the
368
00:20:03,880 --> 00:20:05,920
initial ML spaces that I worked in.
369
00:20:05,920 --> 00:20:08,815
That, or you first build the
model you have some pipeline, and
370
00:20:08,815 --> 00:20:11,875
then you start measuring and see
what matrix you want to measure.
371
00:20:11,875 --> 00:20:14,755
That's not how it really works and
that's not how you should do it.
372
00:20:15,175 --> 00:20:16,285
You start with the matrix.
373
00:20:16,285 --> 00:20:19,675
You start with the steps of where
you are logging, and like you said,
374
00:20:20,215 --> 00:20:23,695
all the things that you said is still
relevant, but there's also how do
375
00:20:23,695 --> 00:20:28,430
you really bring in effectively a
human in the loop for multiple angles
376
00:20:28,825 --> 00:20:30,745
to observe a pattern of output?
377
00:20:30,835 --> 00:20:34,045
Make sure that pattern makes
sense, because if you really
378
00:20:34,045 --> 00:20:35,125
want to scale a system.
379
00:20:35,515 --> 00:20:39,475
There's no way you can scale it with
like say two or three human in the loop
380
00:20:39,475 --> 00:20:43,285
reviewing everything, say thousands of
document data that you have extracted.
381
00:20:43,555 --> 00:20:47,455
But there will be pattern when you
analyze the extracted data to see are
382
00:20:47,455 --> 00:20:49,825
there specific, just an example, right?
383
00:20:49,825 --> 00:20:54,385
Are there specific indicators that
have seemed to have failed the multiple
384
00:20:54,385 --> 00:20:56,605
times for the same type of documents?
385
00:20:56,605 --> 00:21:00,955
That's the pattern you're looking for so
that you know exactly either your, it's
386
00:21:00,955 --> 00:21:03,415
your model that's failing or maybe the.
387
00:21:03,955 --> 00:21:07,750
Kind of extraction you were doing just
an example, for example, workflow, the,
388
00:21:07,825 --> 00:21:09,265
the kind of extraction you were doing.
389
00:21:09,265 --> 00:21:10,075
Maybe it's failing.
390
00:21:10,495 --> 00:21:10,945
One.
391
00:21:10,945 --> 00:21:15,325
I I can also share a learning that I
had, uh, recently, is that some of the
392
00:21:15,535 --> 00:21:17,230
huge financial documents you will see is.
393
00:21:18,640 --> 00:21:22,360
Highly it, it's text, it's text
driven, but it also has like
394
00:21:22,360 --> 00:21:26,080
infographics tables that looks like
tables with's actually an image.
395
00:21:26,350 --> 00:21:29,500
So how do you really bring together
the insights from a text based,
396
00:21:30,190 --> 00:21:34,990
text-based extraction and from like a
table as well as info, uh, infographic
397
00:21:34,990 --> 00:21:36,280
and make sure they all align.
398
00:21:36,280 --> 00:21:39,250
And the summary of all that
extraction really makes sense
399
00:21:39,250 --> 00:21:40,750
and tells the story correctly.
400
00:21:40,780 --> 00:21:44,560
And there's not any no specific data
that's getting lost or you're not
401
00:21:44,770 --> 00:21:46,660
losing anything in, in, in translation.
402
00:21:46,660 --> 00:21:46,960
So.
403
00:21:47,515 --> 00:21:52,975
Yeah, like at, at every stage where
you are bringing in, um, uh, the
404
00:21:52,975 --> 00:21:56,905
Observability factor, be it the human in
the loop, finding the matter, uh, the,
405
00:21:56,905 --> 00:22:01,975
the pattern, and making sure the end
to end story that you are building or
406
00:22:01,975 --> 00:22:06,685
like the solution that you're building
really aligns with the expected output.
407
00:22:06,805 --> 00:22:10,495
And then you build on top of it
slowly to really add, add more
408
00:22:10,495 --> 00:22:12,475
color to it, but start simple.
409
00:22:12,595 --> 00:22:12,835
Yeah.
410
00:22:13,390 --> 00:22:14,140
Yeah, makes sense.
411
00:22:14,410 --> 00:22:17,635
Especially as you kind of described
this agentic workflow, which mixes.
412
00:22:18,280 --> 00:22:20,380
You know, structured
and unstructured, right?
413
00:22:20,590 --> 00:22:20,770
Mm-hmm.
414
00:22:20,770 --> 00:22:24,550
And, and, uh, when you are in
a financial domain, uh, uh, you
415
00:22:24,550 --> 00:22:26,080
cannot hallucinate on numbers.
416
00:22:26,140 --> 00:22:30,580
You know, you cannot pad an extra
zero for a mortgage rate or, or, or,
417
00:22:30,580 --> 00:22:31,990
or like, uh, or some sort of, uh,
418
00:22:32,200 --> 00:22:32,950
intolerance Yeah.
419
00:22:33,160 --> 00:22:33,535
Tolerance there.
420
00:22:34,210 --> 00:22:36,490
Uh, how, how, how do you deal with it?
421
00:22:36,490 --> 00:22:40,300
Because this is a, actually a big problem
that a lot of us are facing, right?
422
00:22:40,300 --> 00:22:41,980
There's this non-deterministic beast.
423
00:22:42,205 --> 00:22:45,985
And you're trying to control and, and
put a deterministic layer on top of it.
424
00:22:46,075 --> 00:22:50,575
Um, how do we, how do we make sure
that this has hallucinations or these
425
00:22:50,575 --> 00:22:52,315
things, you know, monitored and observed?
426
00:22:52,765 --> 00:22:53,125
Yeah.
427
00:22:53,545 --> 00:22:56,875
So that's why I said, right,
like at least for the financial
428
00:22:56,875 --> 00:22:58,135
use cases, I do not think that.
429
00:22:58,990 --> 00:23:04,900
The current LLMs or like the large
language models way they way they are
430
00:23:04,900 --> 00:23:08,380
built and how they output, we all know
they are like all predictive token based.
431
00:23:08,380 --> 00:23:11,680
They're not really understanding the
quantitative numbers and that's why
432
00:23:11,890 --> 00:23:15,880
we've seen, I'm sure you have seen
those, uh, funny LinkedIn posts about
433
00:23:15,880 --> 00:23:19,810
is 9.9 bigger or 9.11 bigger and
then you see all the wrong answers.
434
00:23:19,840 --> 00:23:20,830
That's why it's happening.
435
00:23:20,980 --> 00:23:21,070
Right.
436
00:23:21,070 --> 00:23:22,600
But I think the models are getting.
437
00:23:23,015 --> 00:23:24,295
Better and better as.
438
00:23:24,435 --> 00:23:28,135
As we are, uh, uh, making a little
bit of infrastructure change and
439
00:23:28,135 --> 00:23:30,055
that this is where the, the business.
440
00:23:30,625 --> 00:23:33,205
Way of building a
solution comes into play.
441
00:23:33,235 --> 00:23:34,615
This is your system design.
442
00:23:34,705 --> 00:23:38,575
It's not necessary that you really
apply LLMs for all your steps.
443
00:23:38,905 --> 00:23:39,745
Yes, of course.
444
00:23:39,745 --> 00:23:43,975
For things like that's highly,
uh, uh, time consuming, like
445
00:23:43,975 --> 00:23:45,475
really extraction of the data.
446
00:23:45,475 --> 00:23:47,275
You do a first pass with the LLMs.
447
00:23:47,275 --> 00:23:50,455
You have your own predictive models
that I'm sure all companies have
448
00:23:50,455 --> 00:23:53,965
built, especially if they're a
legacy company in the space that
449
00:23:53,965 --> 00:23:55,915
was doing this work prior to.
450
00:23:56,245 --> 00:23:57,835
The LLMs came into the scene, right?
451
00:23:57,835 --> 00:24:01,915
So why, why are we like moving
completely outta that space and not
452
00:24:01,915 --> 00:24:06,895
really building a, a hybrid space or
a hybrid model that's taking advantage
453
00:24:06,895 --> 00:24:08,845
of what we have been doing for so long?
454
00:24:09,115 --> 00:24:10,975
Use that as our learning curve.
455
00:24:11,035 --> 00:24:15,720
Uh, use that as our, I would say a
training material, or I would say some.
456
00:24:16,375 --> 00:24:19,765
Maybe if I can say it in this way,
maybe a few short based learning
457
00:24:19,765 --> 00:24:22,645
methodology from what we were
doing with our predictive models.
458
00:24:23,605 --> 00:24:23,665
Yeah.
459
00:24:23,665 --> 00:24:26,635
And help leverage and
ground our LM outputs.
460
00:24:26,905 --> 00:24:29,125
Again, there is no one
easy way of doing it.
461
00:24:29,125 --> 00:24:32,845
It's, yeah, a lot of trial and
error, figuring out where we are, our
462
00:24:32,845 --> 00:24:35,905
systems are failing, which indicators
are too complicated for the system
463
00:24:35,905 --> 00:24:37,855
to handle, and then bringing in.
464
00:24:38,140 --> 00:24:42,450
Either SMEs or our previous generation
of models to help us guide through that.
465
00:24:43,140 --> 00:24:43,650
So yeah,
466
00:24:43,890 --> 00:24:47,670
in other words, you're saying, uh,
the classical machine learning models
467
00:24:47,670 --> 00:24:52,170
still need to exist and you need to
layer on an agent, skin on them and
468
00:24:52,170 --> 00:24:55,890
try to ground your generative AI
outputs on the predictive, you know,
469
00:24:55,890 --> 00:24:57,390
what predictive models are suggesting.
470
00:24:58,015 --> 00:25:01,915
There's a lot of value with sound
predictive models that I've seen
471
00:25:01,915 --> 00:25:03,565
the financial industry usually work.
472
00:25:03,625 --> 00:25:06,355
There's a lot of work going on,
and especially now with the power
473
00:25:06,355 --> 00:25:09,865
of LLMs, uh, knowledge graphs have
again, come into the space, right?
474
00:25:09,865 --> 00:25:12,550
Like it's getting a lot of more
traction because now it's easy to.
475
00:25:13,345 --> 00:25:18,265
Easy to stand it up and maintain with
the help of the LLMs, uh, compared to
476
00:25:18,265 --> 00:25:22,795
prior to this as well as a lot of work is
happening on the causal AI side, which is
477
00:25:22,795 --> 00:25:25,435
also a domain I'm highly interested in.
478
00:25:25,435 --> 00:25:29,185
I'm trying to like, find and do as
much work as possible in there is that,
479
00:25:29,185 --> 00:25:34,225
how do you really ground your, um,
non-deterministic outputs from LLMs
480
00:25:34,225 --> 00:25:35,875
with like the causal analysis that.
481
00:25:36,535 --> 00:25:40,645
I am sure you have already done, or your
econometrics team or your statistical
482
00:25:40,645 --> 00:25:42,205
team have been doing prior to this.
483
00:25:42,235 --> 00:25:47,905
Uh, and then utilize that to gauge
your level of correctness, if I can.
484
00:25:47,965 --> 00:25:51,505
Uh, it's not accurate, but correctness,
how correct are you and where
485
00:25:51,505 --> 00:25:54,505
are the areas of um, uh, yeah.
486
00:25:54,595 --> 00:25:55,590
Most of the gaps.
487
00:25:56,905 --> 00:25:57,475
Absolutely.
488
00:25:57,835 --> 00:26:00,535
So there's a question from the audience
I'd like to take at this point.
489
00:26:00,805 --> 00:26:00,925
Mm-hmm.
490
00:26:01,165 --> 00:26:04,825
Um, someone is asking, our agent
system works great in dev mode, but we
491
00:26:04,825 --> 00:26:06,445
keep hearing production is different.
492
00:26:06,805 --> 00:26:10,555
What specific failure mode should
we be preparing for so that, you
493
00:26:10,555 --> 00:26:11,870
know, te our testing isn't catching?
494
00:26:13,755 --> 00:26:15,625
Okay, so it's basically,
could you repeat that?
495
00:26:15,625 --> 00:26:16,495
So it's basically saying the,
496
00:26:16,495 --> 00:26:19,315
Basically it's an, you know, it's the
gen system, you know, you evaluate
497
00:26:19,315 --> 00:26:22,405
it, it works fine on your test cases,
but when you get into production Yeah.
498
00:26:22,735 --> 00:26:27,145
Then you in and all kinds of noisy inputs,
it seems to be having reliability issues.
499
00:26:27,520 --> 00:26:28,180
Absolutely.
500
00:26:28,330 --> 00:26:30,190
So this, this, this happens, right?
501
00:26:30,190 --> 00:26:35,020
Like when either we are not working,
we meaning the engineers and the
502
00:26:35,020 --> 00:26:37,510
technologists or the developers are
not working close enough with the
503
00:26:37,510 --> 00:26:41,260
business to really understand their
edge cases or what their clients are
504
00:26:41,260 --> 00:26:45,160
looking for, what kind of questions
or kinds of dataset that might come in
505
00:26:45,160 --> 00:26:46,660
that we have not tested in the system.
506
00:26:46,660 --> 00:26:47,530
That's number one.
507
00:26:47,800 --> 00:26:51,550
Like I and I have seen this
multiple times with my prior
508
00:26:51,550 --> 00:26:52,960
organizations as well as that.
509
00:26:53,620 --> 00:26:54,700
I am guilty of that too.
510
00:26:54,700 --> 00:26:57,670
Like I said in the beginning, to really
start building the moment we see a new
511
00:26:57,670 --> 00:27:03,250
architecture or a new process and say
that, okay, look, I have 25, uh, tested
512
00:27:03,250 --> 00:27:05,050
data sets or like, even a hundred.
513
00:27:05,050 --> 00:27:06,580
It works absolutely amazing.
514
00:27:06,580 --> 00:27:08,320
Let's put it into a, a prod.
515
00:27:08,840 --> 00:27:12,230
Before that, did you put it into
qa, open it up to your beta testers?
516
00:27:12,260 --> 00:27:16,580
Did you get like a handful of sample
people to really push it to its edge
517
00:27:16,580 --> 00:27:18,230
and figure out where it's breaking?
518
00:27:18,350 --> 00:27:21,650
And with all those Observability
checkpoint that we talked about, that
519
00:27:21,650 --> 00:27:25,790
is how you really catch what, what your
product is missing that maybe your PM has
520
00:27:25,970 --> 00:27:30,170
think about before, or your developers
haven't considered in the edge cases and
521
00:27:30,260 --> 00:27:32,510
you really refine that with your PET test.
522
00:27:32,510 --> 00:27:35,000
I'm a hundred percent sure,
even if you're going down the.
523
00:27:35,355 --> 00:27:38,925
Predictive analytics route, there
is always something that you miss in
524
00:27:38,925 --> 00:27:43,995
production, and this is how you really
fine tune your, or tune iteratively your,
525
00:27:44,205 --> 00:27:47,445
uh, application, software application
or whatever application you're building.
526
00:27:47,685 --> 00:27:50,395
So yeah, never release
the first into prod.
527
00:27:50,415 --> 00:27:53,535
Make sure you're releasing in
QA, dev have open it up with
528
00:27:53,535 --> 00:27:55,745
beta testers, uh, test them.
529
00:27:56,035 --> 00:27:57,795
Ask them to test it and push them.
530
00:27:58,400 --> 00:28:02,870
Push the application to its max and
that's how you really, really know
531
00:28:02,870 --> 00:28:07,070
about your application more and, uh,
get, get your stakeholders buy-in
532
00:28:07,070 --> 00:28:08,240
to really support you in that.
533
00:28:09,260 --> 00:28:09,590
Yeah.
534
00:28:09,920 --> 00:28:12,200
So that sort of brings
to a meta point, right?
535
00:28:12,510 --> 00:28:16,310
Where do you see the biggest gap today
between, you know, current agent system
536
00:28:16,310 --> 00:28:20,470
capabilities as you, you know, build on
top of these orchestration frameworks?
537
00:28:21,210 --> 00:28:25,520
There and then what actually enterprises
need for a reli reliable production,
538
00:28:25,520 --> 00:28:27,200
deployment and MA and maintenance.
539
00:28:28,180 --> 00:28:28,720
Yeah.
540
00:28:29,110 --> 00:28:30,520
So multifold, right?
541
00:28:30,520 --> 00:28:34,480
Like I, I definitely feel,
like I said, I, I, the, the
542
00:28:34,480 --> 00:28:35,980
frameworks are changing every day.
543
00:28:35,980 --> 00:28:41,830
So, so having one framework or access
to one model shouldn't be your moat.
544
00:28:42,610 --> 00:28:45,730
This is the word everybody keeps
throwing around the modeling model
545
00:28:45,730 --> 00:28:47,440
and systems are not your moat.
546
00:28:47,530 --> 00:28:47,890
The.
547
00:28:48,380 --> 00:28:54,080
Application you are building based
on your business' input, the ROI
548
00:28:54,080 --> 00:28:58,610
you're trying to get, the problem
that you're trying to solve, uh, the
549
00:28:58,700 --> 00:29:02,570
matrix that you have designed and
the process workflow that you have.
550
00:29:02,900 --> 00:29:06,560
All that comes together and this kind
of very currently ties back to that.
551
00:29:06,790 --> 00:29:10,360
Great article you shared, uh, Krishna
about the compounding AI systems.
552
00:29:10,750 --> 00:29:14,620
Like I said, it's no longer that
model's responsibility to get you the
553
00:29:14,620 --> 00:29:18,280
right answer and you cannot blame,
oh, the LLMs are hallucinating.
554
00:29:18,280 --> 00:29:20,950
Of course they're hallucinating
because they are built on like.
555
00:29:21,355 --> 00:29:25,945
Entire internet and it doesn't know the
specific requirements for your task if
556
00:29:25,945 --> 00:29:28,615
you do not give it the right directions.
557
00:29:28,645 --> 00:29:34,405
The, uh, the config files full of
system prompts that has very specific
558
00:29:34,405 --> 00:29:38,455
nuanced guidelines, a structured
workflow to follow a structured
559
00:29:38,455 --> 00:29:39,805
workflow, meaning like a set of.
560
00:29:40,045 --> 00:29:43,915
Agents, if you're building an agent
system that follows or have a scope
561
00:29:44,155 --> 00:29:47,155
defined for that agent so that
they're not, they don't have the
562
00:29:47,155 --> 00:29:49,405
tendency to go, uh, out of the box.
563
00:29:49,555 --> 00:29:52,585
One funny thing I actually recently read
is that somebody asked an agent that
564
00:29:52,825 --> 00:29:57,445
help me with this, and then it started
garling out like thousands of lines of
565
00:29:57,445 --> 00:30:00,595
code because that's the way the agent
was trying to help the user, right?
566
00:30:00,595 --> 00:30:06,535
But if you tell me that, help me do 1,
2, 3, by using 5, 6, 7, these tools,
567
00:30:06,925 --> 00:30:08,845
that is how you really build up.
568
00:30:09,465 --> 00:30:14,085
More aware and context aware, like this,
a context aware system that you have a
569
00:30:14,085 --> 00:30:18,495
better potential of evaluating, handling,
and observing, rather than just saying
570
00:30:18,495 --> 00:30:20,835
that, Hey, help me to solve this problem.
571
00:30:20,835 --> 00:30:22,185
So it'll help it like it can.
572
00:30:22,935 --> 00:30:23,445
Absolutely.
573
00:30:23,835 --> 00:30:27,855
Now, now we have all these, uh, agentic
frameworks that it, they're being
574
00:30:27,855 --> 00:30:29,355
catered to business users, right?
575
00:30:29,355 --> 00:30:32,685
So where, where you can go and
make these questions and, you
576
00:30:32,685 --> 00:30:33,885
know, shoot yourself in the foot.
577
00:30:34,245 --> 00:30:34,635
Yeah.
578
00:30:34,810 --> 00:30:34,870
Yeah.
579
00:30:35,890 --> 00:30:36,190
Awesome.
580
00:30:36,640 --> 00:30:38,440
So let's take another
audience question here.
581
00:30:38,500 --> 00:30:42,670
Uh, so someone was asking our agent
systems generate tons of logs, like
582
00:30:42,670 --> 00:30:47,320
reasoning, API calls memory access
planning steps when something fails, how
583
00:30:47,320 --> 00:30:50,980
do you trace through all of that to follow
the, you know, follow the logic chain?
584
00:30:50,980 --> 00:30:52,300
How do you do root cause analysis?
585
00:30:53,090 --> 00:30:54,380
Yeah, yeah, absolutely.
586
00:30:54,380 --> 00:30:57,620
And they can return if you write
the systems the right way, you
587
00:30:57,620 --> 00:31:01,160
can return like from, uh, some
of my, uh, reflection agents.
588
00:31:01,160 --> 00:31:05,030
I have returned the context that
went into the final reflection chain.
589
00:31:05,030 --> 00:31:09,500
I, uh, returned the chain of
thought sometimes a chain of draft.
590
00:31:10,260 --> 00:31:13,890
Chain of draft is nothing but like a
simplified, lesser number of tokens to
591
00:31:13,890 --> 00:31:16,950
use for chain of thought so that you're
not spending too much on like, you
592
00:31:16,950 --> 00:31:19,770
know, uh, on those, on those processes.
593
00:31:19,860 --> 00:31:24,030
Then it also tracks like exact,
um, uh, match points depending
594
00:31:24,030 --> 00:31:25,620
on if you have specific matrix.
595
00:31:25,620 --> 00:31:28,440
There is like if it's contextual,
relevancy and all that
596
00:31:28,500 --> 00:31:30,420
industry standard type stuff.
597
00:31:30,840 --> 00:31:32,900
Uh, and then you can
also add custom metrics.
598
00:31:32,900 --> 00:31:38,440
So I, I feel like this is where your
creativity should come in as a developer.
599
00:31:38,440 --> 00:31:42,250
That what you really need, and that's
why I said start building slowly,
600
00:31:42,430 --> 00:31:44,320
have three, four basic components.
601
00:31:44,320 --> 00:31:48,760
See where you're missing, go back, refine
your evaluation criteria, and then make
602
00:31:48,760 --> 00:31:53,560
sure that you're continuously observing
or you maybe say continuously monitoring.
603
00:31:53,800 --> 00:31:59,700
And that's how you really build a good
quality it, developing solution, uh,
604
00:31:59,700 --> 00:32:03,320
rather than trying to build it all at once
without like, uh, much thought into it.
605
00:32:04,080 --> 00:32:05,640
Yeah, that's actually brings a question.
606
00:32:05,640 --> 00:32:06,990
You mentioned custom metrics, right?
607
00:32:06,990 --> 00:32:12,780
So like, you know, there is no
out-of-the-box way to measure.
608
00:32:13,165 --> 00:32:17,545
These things because like every ap, every
application that you're building is,
609
00:32:17,545 --> 00:32:23,485
is sort of slightly different from the
use case and so, you know, you, you may
610
00:32:23,485 --> 00:32:28,675
have to write, you know, tests or custom
evals to measure those things, right?
611
00:32:28,675 --> 00:32:28,765
Mm-hmm.
612
00:32:29,065 --> 00:32:32,185
So what would be that
development process look like?
613
00:32:32,185 --> 00:32:35,035
For example, let's say you, let's
say you're trying to build a customer
614
00:32:35,035 --> 00:32:37,315
support agent that making a call.
615
00:32:37,315 --> 00:32:37,405
Mm-hmm.
616
00:32:38,120 --> 00:32:41,240
Ticketing platform and
summarizing the, the activity.
617
00:32:41,270 --> 00:32:41,360
Mm-hmm.
618
00:32:41,600 --> 00:32:45,680
What are some of the basic stuff when
a developer needs to do to get, just
619
00:32:45,680 --> 00:32:46,945
get that going in the first place?
620
00:32:47,635 --> 00:32:47,995
Yeah.
621
00:32:48,205 --> 00:32:51,745
So yeah, like I said, like if you're
using a agent framework and if you're
622
00:32:51,745 --> 00:32:54,925
using all the components like tool,
calling, memory, make sure you're
623
00:32:54,925 --> 00:32:56,395
logging each of those components.
624
00:32:56,395 --> 00:32:57,175
That's baseline.
625
00:32:57,175 --> 00:33:00,415
That's industry specific, or
like by industry specific,
626
00:33:00,415 --> 00:33:01,765
I mean industry standard.
627
00:33:02,035 --> 00:33:04,105
That has nothing to do with
your specific use case.
628
00:33:04,525 --> 00:33:08,005
When you start building your use
case with, say, for a customer
629
00:33:08,005 --> 00:33:11,215
agent that you said, I'm sure you're
working with somebody from the
630
00:33:11,215 --> 00:33:16,195
marketing or the sales or whoever,
which departments are the, the.
631
00:33:16,260 --> 00:33:19,980
Partnership between business
side components and developers
632
00:33:19,980 --> 00:33:21,960
have never been this important.
633
00:33:22,140 --> 00:33:22,350
Mm-hmm.
634
00:33:22,350 --> 00:33:27,090
Uh, now that we are into, like really
going into this, uh, open sea of
635
00:33:27,150 --> 00:33:31,350
non-determinism, and there could be
so much, there could be so much, um,
636
00:33:31,380 --> 00:33:36,210
negative or noisy feedback that for me
as a developer without much business
637
00:33:36,210 --> 00:33:41,800
context, I can, I can maybe, I can
maybe feed out the, the the junk from
638
00:33:41,800 --> 00:33:46,660
an initial view, but what could be junk
for me may not be for the business side.
639
00:33:46,660 --> 00:33:50,560
That could have, they could have like
some internal insights that I'm missing.
640
00:33:50,800 --> 00:33:55,450
So working with partnership with business
is your best bet to start from the start.
641
00:33:55,600 --> 00:33:59,560
And that's why I feel, um, everybody
has spoken to in the, uh, in the
642
00:33:59,560 --> 00:34:02,980
finance industry style, who is
trying to build care systems first
643
00:34:02,980 --> 00:34:04,480
talks about that we have to make.
644
00:34:05,215 --> 00:34:06,895
Our, get our stakeholders buy in.
645
00:34:06,895 --> 00:34:09,055
And your stakeholder is
your business counterpart.
646
00:34:09,295 --> 00:34:13,105
So they need to believe in that vision
of what you're trying to build and
647
00:34:13,105 --> 00:34:14,755
how it'll make their life easier.
648
00:34:14,995 --> 00:34:16,225
That's how you bring them in.
649
00:34:16,525 --> 00:34:18,265
Describe the entire process.
650
00:34:18,265 --> 00:34:22,885
Make them understand what the, the
key risks are and help get their help
651
00:34:22,885 --> 00:34:24,445
to really define those key matrix.
652
00:34:24,445 --> 00:34:27,865
There's like literally no other
way that you can, uh, define this
653
00:34:27,865 --> 00:34:32,000
matrix by yourself being a developer,
um, without much business context.
654
00:34:32,860 --> 00:34:33,460
Absolutely.
655
00:34:33,880 --> 00:34:37,540
And so, uh, the question is like, you
know, the world has moved in the, in
656
00:34:37,540 --> 00:34:41,740
the last five, six years from MLOps
to LLMOps to now AgentOps, right?
657
00:34:41,980 --> 00:34:42,280
Yeah.
658
00:34:42,460 --> 00:34:44,410
What are some of the, what are some of the
659
00:34:44,795 --> 00:34:45,495
AgentOps okay.
660
00:34:45,635 --> 00:34:45,855
Yes.
661
00:34:47,020 --> 00:34:49,570
What are some of the differences
or commonalities that
662
00:34:49,570 --> 00:34:50,890
you've seen in these worlds?
663
00:34:50,890 --> 00:34:51,735
You know, uh, mm-hmm.
664
00:34:52,115 --> 00:34:53,950
You know, what are some of
the learnings that you could
665
00:34:53,950 --> 00:34:55,750
probably derive from your LLMOps?
666
00:34:56,125 --> 00:34:59,015
Uh, systems to like now the agentOps.
667
00:34:59,035 --> 00:35:00,295
And what are some of the differences?
668
00:35:00,295 --> 00:35:02,280
That as well, when you do,
we need to do differently.
669
00:35:03,175 --> 00:35:03,475
Yeah.
670
00:35:03,475 --> 00:35:04,225
Yeah, absolutely.
671
00:35:04,375 --> 00:35:09,025
I mean, some concepts still still
have the exact same amount of value.
672
00:35:09,025 --> 00:35:13,345
Like are your, the basic first
level is, is your system useful?
673
00:35:13,435 --> 00:35:17,035
Is it useful in the terms of being
accurate, in the terms of being,
674
00:35:17,065 --> 00:35:20,095
responding to what the user is
asking and not just giving it
675
00:35:20,095 --> 00:35:21,685
like five different links that.
676
00:35:21,955 --> 00:35:23,845
The question didn't even ask for.
677
00:35:23,845 --> 00:35:28,195
So yes, relevancy, accuracy, making
sure the system is not breaking down
678
00:35:28,195 --> 00:35:33,715
or there's no up downtime is your
baseline matrix that rolls over for
679
00:35:33,745 --> 00:35:37,825
each of these components that you
said right now with more agent side.
680
00:35:37,825 --> 00:35:41,305
And there's actually a, a, a
great paper and I can, I can share
681
00:35:41,305 --> 00:35:45,055
the link in the chat, uh, that
recently I was, I was reading this
682
00:35:45,055 --> 00:35:47,245
yesterday night and I found it.
683
00:35:47,395 --> 00:35:50,815
Really, uh, interesting because it aligns.
684
00:35:50,815 --> 00:35:53,455
I'm not sure how I share
a share in the chat.
685
00:35:53,545 --> 00:35:54,475
Maybe I can quickly
686
00:35:55,345 --> 00:35:55,705
yeah.
687
00:35:56,005 --> 00:35:56,455
Share my,
688
00:35:56,995 --> 00:35:59,755
yeah, there should be a way our
team can enable you with that.
689
00:36:00,215 --> 00:36:00,735
No worries.
690
00:36:00,735 --> 00:36:01,975
I can quickly share my screen.
691
00:36:01,975 --> 00:36:06,070
Can you see the, can you see like a paper?
692
00:36:07,210 --> 00:36:07,480
Yep.
693
00:36:08,110 --> 00:36:10,180
Okay, so this is actually
a great paper call.
694
00:36:10,180 --> 00:36:12,310
Why do multi-agent llms fail?
695
00:36:12,370 --> 00:36:16,810
And I was just looking through
some, uh, failure categories
696
00:36:16,810 --> 00:36:18,700
that it has identified here.
697
00:36:18,700 --> 00:36:22,810
If you see the biggest one is like,
say system design agent coordination.
698
00:36:22,870 --> 00:36:26,260
The hardest part is like, how do you
really make sure one agent is working
699
00:36:26,260 --> 00:36:29,860
to whether the other agent the right
way as well as task verification.
700
00:36:29,860 --> 00:36:31,660
So yes, premature termination.
701
00:36:31,805 --> 00:36:36,065
Making sure there is no incomplete
verification or incorrect verification
702
00:36:36,065 --> 00:36:42,155
that these are like very new terminologies
or new ways of thinking for your system.
703
00:36:42,395 --> 00:36:46,985
But, but there's no way you can like
really avoid it because if you're building
704
00:36:46,985 --> 00:36:51,545
a multi-agent, different orchestration
processes, these has to be taken care of
705
00:36:51,575 --> 00:36:55,655
in like one of those chains of loggings
that you're doing and thinking through it.
706
00:36:56,075 --> 00:36:56,600
So yeah.
707
00:36:57,295 --> 00:37:03,535
Very newer, new, I would say newer,
uh, uh, domain of like, how do we
708
00:37:03,535 --> 00:37:07,345
really think about the systems, but the
baseline still remains the same, right?
709
00:37:07,345 --> 00:37:08,245
Like, is it accurate?
710
00:37:08,245 --> 00:37:10,045
Is it helping your end of business user?
711
00:37:10,225 --> 00:37:12,540
Is it giving you the right kind
of answer that you're looking
712
00:37:13,825 --> 00:37:14,185
Absolutely.
713
00:37:14,575 --> 00:37:17,575
So that just tells you that there's
no, is just no black or white.
714
00:37:17,575 --> 00:37:21,385
There's actually a big gray zone
here because the in incompleteness
715
00:37:21,385 --> 00:37:24,085
actually is like a, you know, gets
you into this big gray zone of.
716
00:37:24,400 --> 00:37:26,470
Uh, how to evaluate these systems.
717
00:37:26,890 --> 00:37:28,930
That's amazing That's really cool.
718
00:37:29,050 --> 00:37:33,940
So I guess like, you know, from a, you
know, like let's say from an SDLC of an
719
00:37:33,940 --> 00:37:38,860
agent development perspective, you want
to sort of just give a simple recipe for
720
00:37:38,860 --> 00:37:41,080
a new AI team starting on this journey.
721
00:37:41,080 --> 00:37:42,490
Like maybe five, six steps.
722
00:37:42,490 --> 00:37:46,720
Just like follow this recipe uhhuh, to
deploy your first few agent DEC caps.
723
00:37:46,720 --> 00:37:47,165
What would that be?
724
00:37:48,850 --> 00:37:49,660
Very interesting.
725
00:37:50,590 --> 00:37:54,820
Again, that recipe can be made in so many
different ways, but if you ask for, like,
726
00:37:54,820 --> 00:37:57,370
say, the top three biggest components.
727
00:37:57,550 --> 00:38:00,670
First component is like,
what is the expected output?
728
00:38:00,670 --> 00:38:05,590
Like what, what is the user, user problem
you're solving and the expected output.
729
00:38:05,590 --> 00:38:09,310
From there, you kind of work back
to see what you have in your system.
730
00:38:09,310 --> 00:38:09,790
So far.
731
00:38:09,790 --> 00:38:10,690
Is there data gap?
732
00:38:11,050 --> 00:38:13,870
Uh, if there is data gap, how
do you fulfill that data gap?
733
00:38:13,960 --> 00:38:14,470
If there is.
734
00:38:15,400 --> 00:38:20,530
Process gap that is easy to, I think
now with all the frameworks is it's easy
735
00:38:20,530 --> 00:38:22,125
to fill in the process gap with like.
736
00:38:22,795 --> 00:38:26,185
A lot of open source, like if you
wanna build a, uh, workflow and
737
00:38:26,185 --> 00:38:29,155
chain, a couple different, uh, tools
and models you can easily do with
738
00:38:29,155 --> 00:38:30,835
something open source like ran graph.
739
00:38:31,195 --> 00:38:32,815
But that's not the point.
740
00:38:33,175 --> 00:38:37,435
When you are putting it all together,
do you know exactly where to put
741
00:38:37,435 --> 00:38:39,445
your checkpoints, who your SMEs are?
742
00:38:39,505 --> 00:38:43,525
Again, all the common themes that I
highlighted comes back here as your
743
00:38:43,525 --> 00:38:48,355
recipe, that you need all those people
in your team or to have your back.
744
00:38:48,595 --> 00:38:49,270
Otherwise, if you're.
745
00:38:49,860 --> 00:38:54,030
Building in a silo, I can
guarantee you the, the, the
746
00:38:54,030 --> 00:38:55,920
product is never going to take off.
747
00:38:55,920 --> 00:38:59,250
And there you'll see like no
adoption because everybody
748
00:38:59,250 --> 00:38:59,855
will be like, Hey, did you.
749
00:39:00,445 --> 00:39:01,435
Check in with me.
750
00:39:01,615 --> 00:39:03,265
This is not what I really wanted.
751
00:39:03,505 --> 00:39:05,125
So do your market fit?
752
00:39:05,185 --> 00:39:07,585
Like they say, uh, analysis really well.
753
00:39:07,855 --> 00:39:11,155
Make sure that you're really solving
a problem that is a bottleneck and
754
00:39:11,155 --> 00:39:14,485
not something that you just want
to build for uh, and use the right,
755
00:39:14,485 --> 00:39:17,130
defining the problem, uh, is
the most important, right?
756
00:39:17,665 --> 00:39:21,745
Find the problems and the matrix that
you are so that you can map at the end
757
00:39:21,745 --> 00:39:26,425
of the day, um, before you get started,
and then you can keep hydrating on it.
758
00:39:26,425 --> 00:39:29,935
And then, um, make it more, I
would say defined and colorful.
759
00:39:30,460 --> 00:39:31,120
Right, right.
760
00:39:31,120 --> 00:39:34,300
And then include all the
stakeholders That's awesome.
761
00:39:34,540 --> 00:39:39,370
So I guess, uh, you know, let's actually
take a few, um, audience question.
762
00:39:39,400 --> 00:39:41,440
Um, there's one interesting question.
763
00:39:41,830 --> 00:39:45,430
In your experience, have you seen
a use case that made you go, this
764
00:39:45,430 --> 00:39:49,510
is awful for agents looking for
a framework to think about what
765
00:39:49,510 --> 00:39:51,050
makes a good versus bad use case?
766
00:39:53,950 --> 00:39:54,565
Very interesting.
767
00:39:54,985 --> 00:39:55,765
So I actually.
768
00:39:57,355 --> 00:39:58,015
Very interesting.
769
00:39:58,015 --> 00:40:04,645
I actually did a, uh, um, uh, read a
recent paper about, uh, how agents,
770
00:40:04,675 --> 00:40:08,545
human agents, and, you know, gen
AI or AI agents, whatever, will
771
00:40:08,545 --> 00:40:11,845
come together to solve different
kinds of, you know, problems.
772
00:40:12,205 --> 00:40:18,025
There's this concept of if it's, if
it's low risk or, or the theme is
773
00:40:18,025 --> 00:40:25,960
if, if the, um, variance tolerated
in the output is really, really low.
774
00:40:26,200 --> 00:40:30,100
That's not something you would really
want to build in an autonomous system.
775
00:40:30,310 --> 00:40:34,120
By that I meaning exactly that same
point you said, if you are building
776
00:40:34,120 --> 00:40:39,760
a financial analyst, uh, agent and
it analyzes a trend pattern and gives
777
00:40:39,760 --> 00:40:43,450
you complete garbage output based
on the spreadsheet document that it
778
00:40:43,450 --> 00:40:47,890
evaluated, do you really want to build
an autonomous, uh, agent on top of that?
779
00:40:48,065 --> 00:40:53,195
Maybe the autonomous agent is helping on
the initial stages of data extraction and
780
00:40:53,195 --> 00:40:58,535
putting it into like some templated format
that will help you or any other predictive
781
00:40:58,535 --> 00:41:03,125
kind of process that you have in place
to expedite that process, but end to end.
782
00:41:03,125 --> 00:41:04,955
Is that a good, good, good use case?
783
00:41:04,955 --> 00:41:07,865
I, maybe we are not there yet.
784
00:41:08,015 --> 00:41:10,685
Maybe in two months or three
months or six months, I don't know.
785
00:41:10,685 --> 00:41:13,180
Like maybe we will get
there, but not yet so far.
786
00:41:13,360 --> 00:41:13,780
Mm-hmm.
787
00:41:13,865 --> 00:41:17,705
So you really think through
from the risk as well as.
788
00:41:18,220 --> 00:41:24,940
Uh, tolerance of fault or false tolerance
for your, for your audience, uh, as
789
00:41:24,940 --> 00:41:27,920
your first grounding level that is
that what you, something should do.
790
00:41:28,855 --> 00:41:29,695
That's a good, good point.
791
00:41:29,695 --> 00:41:34,995
You know, variance in terms of being able
to be happy with the Correct output.
792
00:41:35,135 --> 00:41:35,215
Yeah.
793
00:41:35,240 --> 00:41:35,530
Yeah.
794
00:41:35,605 --> 00:41:36,955
It's a q and a procedure.
795
00:41:37,075 --> 00:41:37,495
Sure.
796
00:41:37,495 --> 00:41:40,465
Like if there is a couple of changes
in the line of my q and a, uh,
797
00:41:40,645 --> 00:41:42,535
conversational agent, I really don't mind.
798
00:41:42,535 --> 00:41:47,395
But if there is a factual error or
like a trend analysis error where
799
00:41:47,425 --> 00:41:51,745
it's supposed to show high and it's
showing no low or no changes, that's
800
00:41:51,745 --> 00:41:57,910
catastrophic in terms of reputation
as well as, our entire business line.
801
00:41:58,540 --> 00:41:59,440
Yeah, makes sense.
802
00:41:59,950 --> 00:42:01,570
Uh, another audience question here.
803
00:42:01,630 --> 00:42:07,060
Um, our leadership asks how do we
know agents that won't randomly
804
00:42:07,060 --> 00:42:08,110
break in front of a client?
805
00:42:08,110 --> 00:42:10,990
You know, it's basically they're
building agents and, and we can give
806
00:42:10,990 --> 00:42:12,850
them accuracy metrics like we used to.
807
00:42:13,090 --> 00:42:16,730
How do you convince stakeholders
to trust something unpredictable?
808
00:42:19,360 --> 00:42:22,045
This is, I think I. Industrywide problem.
809
00:42:22,135 --> 00:42:24,565
Uh, you start by educating them.
810
00:42:24,565 --> 00:42:28,825
And this is what, uh, with everything
that I do like outside my day-to-day,
811
00:42:28,825 --> 00:42:32,965
even during my day-to-day work, my entire
goal is that if I'm talking to somebody
812
00:42:32,965 --> 00:42:36,625
from the non-technical side or even from
the technical side who are more from the
813
00:42:36,925 --> 00:42:40,040
predictive analytic side or more from
the software development side, how do I.
814
00:42:40,565 --> 00:42:44,615
Add a little bit of extra information
in any conversation I'm having so
815
00:42:44,615 --> 00:42:46,715
that they have some new perspective.
816
00:42:46,745 --> 00:42:49,655
They don't have to agree with
me, but, so something like that.
817
00:42:49,685 --> 00:42:52,175
Okay, that's a takeaway
for me to go and read up.
818
00:42:52,565 --> 00:42:56,315
Uh, and then I share some materials or
I tell them, what do you think about it?
819
00:42:56,555 --> 00:43:00,215
So start having this conversation
so that your leadership or your.
820
00:43:01,645 --> 00:43:05,545
Other side of the technical or
non-technical stakeholders doesn't feel
821
00:43:05,545 --> 00:43:09,595
like their opinions doesn't matter, and
you are trying to like kind of bulldoze
822
00:43:09,595 --> 00:43:11,155
and build on top of their systems.
823
00:43:11,185 --> 00:43:12,445
That's not how we're gonna.
824
00:43:12,985 --> 00:43:17,365
Make it productive and effective, but it
should be more like really collaborative,
825
00:43:17,395 --> 00:43:21,085
trying to understand where they're
coming from, where their fears are, and
826
00:43:21,085 --> 00:43:23,515
some of those fears are really legit.
827
00:43:23,515 --> 00:43:26,905
So you really don't know if
you can, uh, stop an agent from
828
00:43:26,905 --> 00:43:28,195
crashing during a live demo.
829
00:43:28,195 --> 00:43:31,165
It happens all the time, but it happens
in traditional software as well.
830
00:43:31,165 --> 00:43:34,015
I've seen so many software
demos where the API calls just.
831
00:43:34,555 --> 00:43:37,225
Fails in the middle, or like
there's some connection, error or,
832
00:43:37,525 --> 00:43:39,265
so it's part of the tech world.
833
00:43:39,355 --> 00:43:42,655
It's part of like building, uh,
smaller components and making sure
834
00:43:42,655 --> 00:43:47,365
they're tight, evaluated, and, uh, you
know, kind of gauged the right way.
835
00:43:47,695 --> 00:43:52,825
So yeah, start with maybe friendly,
uh, education and upskilling and
836
00:43:52,825 --> 00:43:55,615
awareness kind of discussions
and, and, and see how that.
837
00:43:56,155 --> 00:43:58,825
How that bonding goes there.
838
00:43:58,825 --> 00:44:00,565
There's really no other way.
839
00:44:00,565 --> 00:44:02,185
Uh, yeah.
840
00:44:02,665 --> 00:44:04,915
So I guess, you know, finally,
to end this conversation, right?
841
00:44:04,915 --> 00:44:10,915
Like as the world is moving from
deterministic software into this agent
842
00:44:10,915 --> 00:44:13,045
tech, non-deterministic software mm-hmm.
843
00:44:13,375 --> 00:44:17,275
How important are the things that we
talked about in evaluation, Observability,
844
00:44:17,275 --> 00:44:18,485
guardrails, you know, where would you.
845
00:44:19,210 --> 00:44:22,660
You know, how do you think, how do
you sort of think about it from an
846
00:44:22,660 --> 00:44:24,340
AI, you know, leadership perspective?
847
00:44:24,340 --> 00:44:25,660
Also from a developer perspective?
848
00:44:26,335 --> 00:44:32,755
Yeah, I mean, at least everything that
I build, I try to make sure that I build
849
00:44:32,755 --> 00:44:37,345
in a way that I can prove if somebody
wants to run that process again, I show
850
00:44:37,345 --> 00:44:41,905
them some lineage of information, even if
there is like some change in ity or like
851
00:44:42,145 --> 00:44:44,635
some change in uh, uh, extraction process.
852
00:44:44,875 --> 00:44:48,295
But the key goal should be to show
that, hey, look, this process works.
853
00:44:48,295 --> 00:44:50,170
It's making, um.
854
00:44:50,905 --> 00:44:53,155
The work that I used to do in like
say, I don't know, just give an
855
00:44:53,155 --> 00:44:57,235
example, five hours now I do in
like one or one and a half and my
856
00:44:57,235 --> 00:44:59,155
productivity boost has been insane.
857
00:44:59,155 --> 00:45:01,645
And look, I have three more
ideas that have come up with that
858
00:45:01,645 --> 00:45:03,115
extra time that I have in hand.
859
00:45:03,445 --> 00:45:04,525
So it's, it's.
860
00:45:04,885 --> 00:45:09,325
It's evaluation is key to doing all of
that, just to make sure that you are
861
00:45:09,325 --> 00:45:13,735
building the right way, thinking in
a more modular structure, and making
862
00:45:13,735 --> 00:45:17,965
sure that each module and here module,
meaning your data pipeline is one module.
863
00:45:18,055 --> 00:45:21,745
The output from that should be somewhere
where you're checking the output quality,
864
00:45:21,745 --> 00:45:23,275
drift direction and all that stuff.
865
00:45:23,575 --> 00:45:27,130
Then your com agent components, or
like the individual business, uh,
866
00:45:27,135 --> 00:45:30,535
uh, components that you have should
have another segmented module.
867
00:45:30,895 --> 00:45:33,865
So yeah, build the, build it.
868
00:45:34,100 --> 00:45:38,390
In a way that you can break it down
easily and then see where the gaps are.
869
00:45:38,390 --> 00:45:43,190
Don't build a monolithic legacy
agent system end to end five agents
870
00:45:43,190 --> 00:45:44,390
calling three different agents.
871
00:45:44,390 --> 00:45:45,830
I don't think we really need that.
872
00:45:46,160 --> 00:45:49,310
Yeah, sometimes people really
complicate stuff just because they
873
00:45:49,310 --> 00:45:50,810
want to build complicated systems.
874
00:45:51,355 --> 00:45:55,405
Sometimes I've seen simple agents with two
different calls that they want to have a
875
00:45:55,405 --> 00:45:59,755
very specific task they want to fulfill,
can do a lot and like save a lot of time.
876
00:45:59,755 --> 00:46:01,495
So think about the 20 80 rule.
877
00:46:01,885 --> 00:46:04,705
Find out the use cases that you want
to prioritize that solves the most
878
00:46:04,705 --> 00:46:08,605
amount of problem, and, and really start
there and make sure you evaluate as you
879
00:46:09,580 --> 00:46:10,060
Awesome.
880
00:46:10,420 --> 00:46:11,260
Thank you so much.
881
00:46:11,290 --> 00:46:14,440
Uh, I think, yeah, that sums
it up, you know, uh, thanks for
882
00:46:14,440 --> 00:46:15,790
spending the time with us today.
883
00:46:15,820 --> 00:46:19,030
Uh, we'd love to see that paper
that you'd, you know, shared.
884
00:46:19,030 --> 00:46:20,470
Maybe you can share a link with us.
885
00:46:20,530 --> 00:46:21,580
Uh, did I do that?
886
00:46:21,580 --> 00:46:24,640
And then we can share it with our,
you know, webinar attendees later.
887
00:46:24,760 --> 00:46:27,310
Um, and, um, yeah, absolutely.
888
00:46:27,310 --> 00:46:28,690
This is a exciting time.
889
00:46:28,690 --> 00:46:32,500
And, and yeah, I think, uh, the
paper is there for everybody to,
890
00:46:33,100 --> 00:46:35,350
I think it just went to post on
the panelists, but yeah, you,
891
00:46:37,420 --> 00:46:37,930
Awesome.
892
00:46:38,440 --> 00:46:38,830
Okay.
893
00:46:38,920 --> 00:46:40,990
Uh, that's it for this week folks.
894
00:46:40,990 --> 00:46:43,420
Uh, thanks for joining us on the AI.
895
00:46:43,420 --> 00:46:47,020
Explain, and, uh, we'll come
back with another great edition
896
00:46:47,020 --> 00:46:48,190
with another great speaker.
897
00:46:48,280 --> 00:46:49,960
Until then, you know, see you.
898
00:46:50,470 --> 00:46:51,010
Thank you.